01 / 09
Private AI
2024 Private AI foundation
Private AI became an operating foundation.
Three on-premises NVIDIA DGX systems established the WHPS private-AI foundation in 2024.
- 3×
- On-premises NVIDIA DGXProgramme-reported count
- 2024
- Foundation establishedStrategic business case timing
- Standing
- Programme-reported count and timing
WHPS private AI operating cutaway
- Systems
- 3× on-premises NVIDIA DGX
- Timing
- 2024 foundation milestone
- Claim boundary
- Conceptual geometry—not installation or configuration evidence
January 2026 strategic business case · reported
The business case records local inference inside the estate.
On-premises NVIDIA DGX running local LLMs in support of a late-2025 production contact-center stack.- 3.6M
- member interactions
- 99.99%
- reported uptime
Documented model operating architecture
Models can change. The control contract does not.
- 01Identity + allowed contextData scope before retrieval
- 02Governed retrievalSource-backed context + citations
- 03Model gatewayRouting policy + replaceable models
- 04Local inferenceProtected runtime boundary
- 05Response controlPrivacy + confidence + policy
- 06Human serviceResolve or escalate
Documented model + retrieval release contract
Every model or retrieval change must earn release.
- 01Register
- 02Evaluate
- 03Independent verify
- 04Human approve
- 05Observe / revoke
Next proof Training, fine-tuning, and a reviewed model inventory are not yet evidenced as operational. Complete the runtime boundary, retrieval-source lineage, evaluation baselines, and training standing before making that claim.
Business-case-reported operating proof · documented architecture and release standard · explanatory geometry · no installation topology or training claim
2025 Operating proof
AI entered the service path.
By late 2025, the business case recorded a production contact-center stack. The milestone was AI entering the service path; the cutaway shows the documented controls and human decision boundary.
The documented path runs from approved channel to bounded response—with human service still in control.
Contact Center AI
One request. One protected route. One accountable resolution.
Documented operating architecture
Protected inference boundary
Human service boundary
- 01 Approved channelVoice or authenticated service entry
- 02 Intent + contextIdentity, conversation state, and allowed context
- 03 Retrieve + inferSource-backed retrieval and local inference
- 04 Response controlPrivacy, confidence, policy, and fail-closed checks
- 05 Human serviceResolve or escalate; document and review Person remains in control
Operating proof
3.6M interactions 99.99% uptime late-2025 production stack
Technology ecosystem
Wipro NVIDIA Cisco Google Cloud
Source boundary: the January 2026 strategic business case states the late-2025 production stack, 3.6M interactions, and 99.99% uptime. Specific workload, open-enrollment, and call-type milestones come from that business case and are not independently validated here. The cutaway is documented target architecture, not independently validated production-flow or installation evidence.
Private AI Model estate
A model earns standing before it earns traffic.
Local inference is business-case reported. The next operating proof is a reviewed record for every model, retriever, evaluation baseline, release decision, and revoke path.
Unknown is an explicit standing—not a blank to fill with a claim.
Model Estate Ledger
One accountable record from workload to revoke.
Standing travels with the asset
-
01
Workload standingContact Center AI · local LLM inferenceBusiness-case reported
-
02
Runtime + routingPrivate boundary · model gateway · replaceable modelsDocumented standard
-
03
Asset identityModel · version · origin · license · approved workloadReviewed inventory open
-
04
Retrieval lineageSources · refresh · permissions · citation pathEvidence required
-
05
Evaluation baselineQuality · privacy · safety · fallback · rollbackEvidence required
-
06
Release + revokeIndependent verifier · human disposition · change recordDocumented standard
Standing boundary: local LLM inference is business-case reported. The lifecycle controls are documented standards. The approved operational model inventory, exact runtime configuration, training or fine-tuning standing, dataset lineage, and evaluation baselines remain open evidence obligations.
WHPS AI SDLC Verification Control Plane
AI-native delivery needs a control system.
Seven phases structure the work. Thirteen gates bind it. Preventive, inline, and gate verifiers can stop it. A named human decides what ships.
MarketLink F001 · governed delivery path
Lifecycle × verifier release instrument
Execution modes Human-led · Hybrid · Agent-led
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
- 09
- 10
- 11
- 12
- 13
01–04 Intake + design assurance 05–08 Build + runtime security 09–12 Access + operations 13 Final risk acceptance
05 / Validate / Test
Active
Validate / Test
Owner Engineering / QA / Compliance
Bounded execution
Every release decision carries its evidence forward.
Documented operating standard
-
01
Intent contract Outcome, constraints, acceptance, standing
-
02
Scoped workspace Bounded context, identity, permissions, and tools
Model gateway Tool gateway
-
03
Independent verifier Different lineage or deterministic check
-
04
Evidence package Result, provenance, disposition, and unresolved gaps
-
05
Named human disposition Approve, hold, remediate, or reject
Segregation of duties The verifier sits outside the scoped workspace. No agent verifies its own output.
Observability tells you what happened. A control plane determines what is allowed to happen. Governance sets the rules; the control plane enforces them. No agent verifies its own output.
-
01Preventive Before The agent cannot. Controls
- PHI data boundary — production stores unreachable from any generation context Algorithmic Blocks
- Workspace credential scope — no standing production or DB2 access is issued Algorithmic Blocks
-
02Inline During Checked while working. Controls
- PHI boundary scan Algorithmic Blocks
- Protected-lane and diff-scope guard Algorithmic Blocks
-
03Gate Before release Nothing ships unverified. Controls
- Intent-versus-implementation review Agentic Flags
- Release authority sign-off Human Approves
Human release authority Architecture, risk, and production approval remain named human decisions.
-
04Continuous After Detects. Does not control. Observes
- Drift and regression watch Algorithmic Flags
Algorithmic and agentic verification are fused — one catches the known, the other catches intent. A layer across the seven phases and thirteen control gates, not an eighth phase. Documented operating standard
Human release disposition Evidence gate held · 0 export-ready · 11 captures pending
Method source · WHPS AI SDLC standard: seven phases, thirteen control gates. MarketLink repository values; traceability coverage is not evidence completion.
MarketLink CMS enrollment journey
Enrollment stays connected from entry to carrier output.
Production-ready. CMS APIs exercised successfully in CMS UAT. Broker and consumer enrollment moves through one controlled journey while the audit and evidence package remains open.
Product outcome Identity, consent, application status, exceptions, and downstream enrollment remain connected.
Journey in one view
One member and application context crosses every handoff.
Production-ready · CMS UAT exercised
Member and application context
-
01
Broker / white-label entryBroker desktop or branded consumer pathway.
-
02
Identity + consentBroker role, consumer proofing, and permission.
-
03
Application + plan selectionHousehold, eligibility, subsidy, and plan choice.
-
04
CMS Hub submit + statusSubmit enrollment and receive application status.
-
05
DMI / SVI resolutionDocuments, notices, policy updates, and follow-up.
-
06
834 / carrier outputEnrollment file, acknowledgment, and handoff.
Status changes, consent defects, and DMI / SVI exceptions return to the same controlled workflow.
Evidence readiness
21 / 21 traceability rows mapped
100% traceability coverage · coverage is not completion.
- 0
- export-ready
- 1
- review candidate
- 11
- pending capture
- 1
- source pending
Current standing Production-ready · CMS UAT exercised Audit and evidence package open
ReconLink Reconciliation intelligence
Two tracks. One acceptance boundary.
Recon Buddy is extending a live phase-one foundation while the ReconLink platform is tested and tuned with Operations. Neither track converts activity into release until evidence closes.
A tentative date is not a release decision. Testing evidence and named acceptance are.
Current governed position
Operating foundation and platform modernization converge at acceptance.
Position as of August 5, 2026
- Phase two testing
- NERD help integration
- Defect correction
- Platform testing
- Weekly Operations feedback
- Model tuning
Shared acceptance plane
- 01Test resultRequired behavior passes
- 02Operations dispositionFeedback is resolved or accepted
- 03Defect boundaryRelease-blocking issues are closed
Named acceptance
Release, hold, or continue the tuning loop.
Testing timeframes govern the release date.
2026 Modernize
Modernization without a big-bang cutover.
GroupLink is client live with payments integrated. Bidirectional IBM CDC is deploying while DB2 remains authoritative; parity evidence gates cutover and workload retirement.
The bridge is the strategy: modernize the path, prove parity, then retire deliberately.
GroupLink coexistence architecture
Modern experience above an explicit system of record
Bidirectional synchronization · deploying
ExperienceGroupLink PortalRole-aware workflows
Service boundaryService APIsVersioned contracts reduce legacy coupling
Modern dataPostgreSQLApplication + read model
IBM CDCBidirectional · deploying
System of recordDB2Authoritative data · batch · payments · controls
Retirement gate remains closed
- Functional parity
- Onboarding
- Batch
- Payments
- Operating readiness
GroupLink is client live with payments integrated. Bidirectional IBM CDC is deploying while DB2 remains authoritative; parity evidence gates cutover. Workload retirement remains future work.
August 2026 Portfolio value
Value follows proof into retirement.
Four product lanes make the transformation tangible. The business case changes only when product outcomes are measured and replaced workloads actually leave the estate.
22
initiatives in the dated transformation register
As of August 5, 2026: 22 initiatives · 11 active delivery · 2 decision-dependent · 8 queued · 2 live / operational. Categories overlap.
Product → proof → retirement
The portfolio earns value in sequence.
Portfolio position + conditional business case
Recognition gates
- 01Measure adoption
- 02Prove operating change
- 03Verify workload retirement
- 04Measure replacement run cost
Operating baseline$55.0M$1.28 PMPM
→
Conditional target$26.5M$0.60 PMPM
$28.5M modeled difference
22 initiatives 11 active delivery 2 decision-dependent 8 queued 2 live / operational
Position as of August 5, 2026 · categories overlap · initiative details are dated January 2026 modeled scenario—not realized savings.
Leadership decision Second-product proof
Prove the system on a second product.
Close MarketLink’s evidence package. Then run ReconLink through the same release contract and measure whether the delivery discipline—not just the team—repeats.
Decision now
Authorize one measured second-product proof.
Use ReconLink as the recommended replication lane; keep independent verifier standing and named
human release authority binding.
Replication protocol
Same contract. Different product. Measured result.
Recommendation · not yet authorized
- Seven phases
- Thirteen gates
- Independent verification
- Named human release
- 01Lead timeIntent to accepted release
- 02ReworkDefects and remediation cycles
- 03EvidenceRequired artifacts complete at decision
- 04AdoptionObserved use and operating outcome
Proof condition
Repeatability exists only if the second lane improves under the same controls.
Result can confirm, refine, or reject the operating hypothesis.
Decision boundary: this is a recommended controlled experiment, not evidence that the method already repeats, a proprietary offering exists, or target economics have been realized.
WHPS × Wipro